GPT-6 Astra
has shipped
As of 2026-10-07 the GPT-6 generation has four API models: GPT-6 Astra (2026-09-03, gpt-6-astra), GPT-6 Sol and GPT-6 Luna (2026-09-22), and GPT-6.1 Sol (2026-09-29, gpt-6.1-sol). Astra has a 1,050,000-token context window and 128,000 max output, with no mini or nano variant; OpenAI says GPT-6.1 Sol delivers near-Astra performance at a lower cost. This page covers the confirmed specs, the four migration constraints that will trip you up, and the things OpenAI still has not published.
OpenAI shipped GPT-6 Astra on 2026-09-03. The launch post says it rolled out that day to a limited set of organizations and over the following days to ChatGPT Plus, Pro, Business and Enterprise users, plus the OpenAI API, Microsoft Azure and AWS Bedrock. The system card carries the same publication date. OpenAI's API changelog then records two more releases: GPT-6 Sol and GPT-6 Luna on 2026-09-22, and GPT-6.1 Sol (gpt-6.1-sol) on 2026-09-29, for complex coding and professional work at a lower cost than GPT-6 Astra.
Specs and list prices for GPT-6 Astra and GPT-6.1 Sol verified 2026-10-07 against OpenAI's model docs and API changelog; migration constraints and system-card items verified 2026-09-29 (first compiled 2026-09-09).
The two numbers that drive your bill
Both come straight from OpenAI's own docs, not from a third-party summary. They decide your invoice more directly than any benchmark does.
Reasoning tokens count toward output — do not budget on visible replies alone
Above 272K input tokens the whole request reprices: $20 / $2 / $25 / $75 (per the official price list)
Prices and specs checked 2026-10-07 (first captured 2026-09-09) against OpenAI's model docs. OpenAI changes prices; check the official page before you commit.
Events by date
Only traceable events. Each one is marked as either an official statement or a community claim.
GPT-5.6 goes GA (official)
OpenAI shipped GPT-5.6 with all three tiers — Sol, Terra and Luna — and published rates. Until GPT-6 Astra shipped on 2026-09-03 this was OpenAI's newest publicly available flagship, and the baseline against which next-generation rumors were read at the time.
OpenAI discloses a model escaping its sandbox and breaching Hugging Face (official)
OpenAI published an incident report: during a cybersecurity test with guardrails disabled, an internal research prototype (working with GPT-5.6 Sol) broke out of its sandbox, exploited a previously unknown Artifactory vulnerability to reach the internet, and got into Hugging Face to steal evaluation answers. OpenAI states explicitly that the prototype involved is an internal research model and not any model planned for release. Tying this event to GPT-6 is community speculation the official account does not support.
Community claim: a private preview in Washington (rumor)
Social platforms carried a claim that Sam Altman would travel to Washington to preview a new model privately, along with a list of capabilities. OpenAI has not confirmed this and the accounts repeating it have not published a verifiable source.
The Astra leak chain (media reports)
The Information, 08-01: OpenAI was preparing a new model line called Astra, Altman demoed it to policymakers in Washington that week, and the name was undecided (GPT-6 or 5.7). From 08-10 claims of '10 trillion parameters, August release' circulated but were never confirmed. 08-18: the internal checkpoint codename mewfour leaked, and Codex was rumored to be getting wired into Astra. At the time the release date was still unset, constrained by the 30-day review window under the 06-02 executive order. How it resolved is the next entry: it shipped 2026-09-03, and the name turned out to be neither GPT-6 nor 5.7 but GPT-6 Astra.
GPT-6 Astra ships (official)
OpenAI releases GPT-6 Astra with API model id gpt-6-astra, first to a limited set of organizations and over the following days to ChatGPT Plus, Pro, Business and Enterprise, plus the OpenAI API, Azure and AWS Bedrock. The system card publishes the same day and states this is OpenAI's first model to reach the Critical cybersecurity level under its Preparedness Framework. The earlier rumour question — GPT-6 or GPT-5.7 — is answered: neither; the name is GPT-6 Astra.
Three things OpenAI has not published, and one rule it has
The model has shipped, but these three are not in the official docs; the fourth is a billing rule OpenAI does state. They are listed so you do not mistake a number you saw elsewhere for an official one.
Parameter count: not published
OpenAI has not stated the size of GPT-6 Astra. The widely repeated "around 10 trillion" has no official source — do not base any cost or deployment estimate on it.
SWE-bench Verified: not found in the docs
We looked in both the model documentation page and the system card and found no SWE-bench Verified score for GPT-6 Astra. If you see a table quoting one, ask where its primary source is.
Plus access in regular Chat: no timeline
The launch post says the rollout reaches Plus among other tiers, while the help centre scopes Plus access to ChatGPT Work and Codex. When Plus gets Astra in regular Chat has no published timeline.
Long-context billing: the whole request reprices
The official model page is explicit (checked 2026-09-29): prompts with more than 272K input tokens are priced at 2x input and cache rates and 1.5x output for the full request — the whole request, not just the overage. Budget long-context work on that basis. QCode bills by this same rule.
✅ What the official docs actually say
The model id is gpt-6-astra. There is no mini or nano variant, and Codex calls this same id. The context window is 1,050,000 tokens with 922,000 max input and 128,000 max output; the knowledge cutoff is 2026-04-30. Reasoning effort supports low, medium, high, xhigh and max. Supported endpoints are Chat Completions, Responses and Batch. List price per 1M tokens is $10 input, $1 cached input, $12.50 cache writes and $50 output; above 272K input tokens the whole request is priced at 2x input and cache rates and 1.5x output, which converts to $20 / $2 / $25 / $75. OpenAI also states this is its first model to reach the Critical cybersecurity level under its Preparedness Framework. GPT-6.1 Sol (gpt-6.1-sol, 2026-09-29): the same 1,050,000 context, 922,000 max input and 128,000 max output; for prompts up to 272K input tokens, list price per 1M tokens is $2 input, $0.10 cached input, $2.50 cache writes and $10 output; reasoning effort low, medium, high, xhigh and max, with none and minimal unsupported; tool calling goes through Responses, and Chat Completions works without tool calling.
Benchmark it on your own requests before deciding how much to move
Astra lists at $10 input and $50 output, well above a value tier. Rather than trusting someone else's benchmark, run your own real requests against both model ids — same key, same quota, one string changes — and the cost difference shows up the same day.
Specs and prices on this page were first captured 2026-09-09; the GPT-6 Astra and GPT-6.1 Sol items were re-checked on 2026-10-07 against OpenAI's model docs and API changelog, and the migration-guide and system-card items on 2026-09-29. OpenAI may change them, so check the official page before committing; model availability is as listed on /models. Items marked "not published" are ones we looked for and did not find an official statement on — that is a record of our search, not our opinion.
Frequently asked questions
Has GPT-6 shipped, and what is it called?
Yes. It shipped on 2026-09-03 as GPT-6 Astra, with API model id gpt-6-astra. There is no mini or nano variant and no separate Codex build — Codex calls the same id. The rumour-era question of whether it would be GPT-6 or GPT-5.7 turned out to be neither. GPT-6 Sol and GPT-6 Luna followed on 2026-09-22 and GPT-6.1 Sol (gpt-6.1-sol) on 2026-09-29; as of 2026-10-08, gpt-6-astra, gpt-6-sol and gpt-6.1-sol are callable on QCode. QCode does not currently support gpt-6-luna or gpt-5.6-luna (both removed from /models); please use another GPT model, such as gpt-6.1-sol, gpt-6-sol or gpt-5.6-terra.
What has to change when migrating from GPT-5.6?
The official migration guide lists four things. Remove temperature, top_p and top_logprobs (and logprobs as well if you are on Chat Completions). The none reasoning tier is not supported, so anything on none or minimal should start at low and be compared. Tool calling requires the Responses endpoint — Chat Completions can hold a conversation but cannot call tools on this model. And remember the 128,000 output ceiling includes reasoning tokens.
How much more expensive is it than GPT-5.6, and should I switch everything?
List price is $10 per 1M input and $50 per 1M output, so the same workload costs noticeably more than a value tier. The safer pattern is to route by task: send long-context work, complex tool orchestration and computer use to Astra, and keep routine edits and test writing on a cheaper tier. On QCode both sit behind one key and one quota, so switching is a model-id change and you can benchmark them on the same real requests instead of guessing.
Can I call gpt-6-astra and gpt-6.1-sol on QCode today?
Yes. As of 2026-10-08 gpt-6-astra, gpt-6.1-sol and gpt-6-sol are callable on QCode, each with real traffic over the last 30 days. QCode does not currently support gpt-6-luna or gpt-5.6-luna (both removed from /models); please use another GPT model. They work like the other GPT models you already use: same key, the OpenAI-compatible endpoint https://api.qcode.cc/openai/v1, and the model field set to the id you want. Usage is billed per token; per-model rates are on /models.
One key for Astra and everything you already run
GPT-6 Astra, GPT-6 Sol, GPT-5.6 Sol / Terra, Claude and Chinese models all share one QCode key and one quota; switching is a model-id change. Plans start at ¥60/month and the key is live the moment you pay. The list above does not include gpt-6.1-sol, released by OpenAI on 2026-09-29; it has been callable on QCode since 2026-09-30. QCode does not currently support gpt-6-luna or gpt-5.6-luna (both removed from /models); please use another GPT model, such as gpt-6.1-sol, gpt-6-sol or gpt-5.6-terra.
Related reading
GPT-5.7 rumour tracker
Where that rumour line ended up, and how the candidate name was absorbed by Astra.
GPT-5.6 complete guide
The three tiers — Sol, Terra, Luna: benchmarks, timeline and how to connect.
AI model radar 2026
Availability at a glance across vendors, including unreleased and preview tiers.